AI Monthly Reports
Notes from our internal AI sessions, with the text in the page.
AI Monthly — May 2025
Two internal sessions: a wearable recorder into transcripts and filing, and the limit of AI layout. Simple pages needed almost no hand-written code; a complex header took longer than doing it by hand.
AI Monthly — June 2025
Three sessions. A coding agent fixed a missing SVG icon before dinner and swapped an unlicensed font across the project in a few hours. The same month counted how communication eats a frontend lead's day.
AI Monthly — July 2025
The internal assistant moved from several Java packages onto n8n, so a logic change no longer meant a full redeploy. Two reading groups compared inventing something with scaling it.
AI Monthly — August 2025
Four sessions: unreleased models on a public leaderboard, Chinese audio summaries, open-weight models running locally, and a proof of concept that masks personal data in logs.
AI Monthly — September 2025
Four sessions. Individual experiments turned into a shared practice. One system-design document was drafted in a few hours by reading the codebase, then finished by hand.
AI Monthly — October 2025
One deep dive. The internal assistant moved from keywords to a visual flow, and a backend engineer used Figma's MCP to build an exhibition animation from the design file. Higher model scores were starting to matter less than the work in front of people.
AI Monthly — November 2025
Four weeks on AI browsers and the ad model, Gemini, and the editor acquisitions, plus in-house trials of a design plugin and of reversing code back into a spec.
AI Monthly — December 2025
Four sessions. Labs chased each other's scores, NotebookLM went from papers to slides, and a CMS prototype started matching page sections to templates. A proposal included a clickable HTML prototype.
AI Monthly — January 2026
Six workshops. The tool platform gained a backend and coworkers published their own utilities; a desktop shortcut made asking a model one keystroke. The old index mentioned OpenClaw, LanceDB, and Ollama; this issue does not.
AI Monthly — February 2026
Three sessions, with a week off for the lunar new year. A local agent moved from a news item onto a machine in the office, and estimates started with a small proof before anyone promised person-days.
AI Monthly — March 2026
Four weeks. Models moved toward operating a computer, and the internal assistant showed up in standups, lookups, and deploys. Similar layouts went from taking months to finishing several in a day. A single speed multiple stays off the page: the before-and-after units do not divide cleanly. Spec turnaround also shortened; that figure is the team's own account.
AI Monthly — April 2026
Three sessions about harnesses rather than model scores: a leaked coding-agent codebase rewritten overnight, a public brokerage-merger incident, and an internal tool platform that agents can publish to.
AI Monthly — May 2026
Four sessions of industry watching: labs renting compute from each other, vendors selling delivery instead of only models, a new class of vulnerability, and the need to govern a growing set of agents.
AI Monthly — June 2026
Three talks: local models and AI as infrastructure, what subsidized subscriptions mean for cost, and how AI engineering moved from prompts to loops.
AI Monthly — July 2026
Four sessions, five topics: how two-factor authentication works, subsidized pricing and on-device models, automating iPhone screenshots with AI, Palantir and the FDE model, and websites being read by AI.
AI Monthly — August 2026
Three sessions: moving from reading AI-written code to verifying it, why customers and delivery matter once models are commodities, and effective tokens and usage pacing.
AI Monthly — September 2026
One session: new models trained largely by the previous generation of AI, and how retrieval-augmented generation (RAG) works, as used in the internal knowledge assistant we built.
AI Monthly — October 2026
Three sessions, four topics: cloud agents, a fog-of-war game and a RAG toolchain, AI time estimates and handoff documents, and vulnerabilities in the age of AI.